[R] R-equivalent Stata command: poisson or quasipoisson?

[Wil M Contreras Arbaje]

Thank you Ben,

 From the article, the purpose of the author's methodology is to  
better handle heteroskedasticity (due, in part, to Jensen's  
inequality). Either way, I'll try both, and see how they compare, as  
I'd like the R estimation to match the Stata one.

Thanks again for your insight,

Cheers,


Wil

On Sep 12, 2010, at 12:58 PM, Ben Bolker wrote:

> Wil M Contreras Arbaje <wil.contreras <at> gmail.com> writes:
>
>>
>> Thanks Bill!
>>
>> Not asking for help with Stata at all, on the contrary: the article
>> mentioned using Stata to fit the model described earlier, and I  
>> wasn't
>> sure how to do the same in R (which is what I've used since college).
>>
>> Thanks again, I'll play around a bit glmRob, see what happens (though
>> it's slightly worrisome that I won't be able to obtain similar
>> results, if only for 'contrast').
>>
>> Cheers,
>>
>> Wil
>
>
>  I find it very hard to tell from Stata's help page, but my best guess
> would be that the previously mentioned Stata command is more or less
> equivalent to R's quasipoisson -- the 'robust' specification seems to
> apply only to the standard error calculation, not to the fitting  
> process.
> What's unclear about 'robust' is that in other (least-squares fitting)
> contexts in Stata, it means 'Huber-White sandwich estimators', i.e.
> estimators that are robust to heteroscedasticity.  I suppose this is
> more general (but also more data-hungry) than the simple expedient of
> scaling the standard errors by a single estimated overdispersion  
> parameter.
>
>  The best thing, of course, would be to try a test case in both
> systems.  Or it seems that
> http://www.stata.com/bookstore/lrm.html (chapter 9) would be helpful.
> (I checked the stata list archives for 'quasipoisson' and found only
> a post from the author ...)
>
>  Somewhat heretically, I prefer polycultures to monocultures; I like
> R for many reasons, but I'm glad that there are other systems out  
> there
> with independent implementations and different sets of advantages
> and drawbacks.
>
>>
>> On Sep 12, 2010, at 12:36 AM, <Bill.Venables <at> csiro.au>  
>> <Bill.Venables
> <at> csiro.au
>>> wrote:
>>
>>> In R, the glm families poisson and quasipoisson will give you the
>>> same estimates.  Their standard errors will (usually) be different,
>>> though, and family = quasipoisson does not give you an AIC (since it
>>> does not maximise a true likelihood; it uses quasi-likelihood
>>> estimation).
>>>
>>> I hope you are not asking this list for help with Stata. We've never
>>> heard of it.  It looks to me, though, that what you are doing below
>>> is fitting a robust poisson glm.  If so, it is something different
>>> again.  There is a package 'robust' which has a glmRob() fitting
>>> function in it that may do something similar, but there is so much
>>> tweaking allowed with robust fits the chance of getting the same
>>> result as with some other system (or even with R if you do it again,
>>> mostly) is effectively zero.
>>>
>>> Tip: use R and forget the others.  It makes life so much easier all
>>> round.
>>>
>>>
>>> -----Original Message-----
>>> From: r-help-bounces <at> r-project.org [mailto:r-help-bounces <at>
> r-project.org
>>> ] On Behalf Of Wil M Contreras Arbaje
>>> Sent: Sunday, 12 September 2010 11:27 AM
>>> To: r-help <at> r-project.org
>>> Subject: [R] R-equivalent Stata command: poisson or quasipoisson?
>>>
>>> Hello R-help,
>>>
>>> According to a research article that covers the topic I'm analyzing,
>>> in Stata, a Poisson pseudo-maximum-likelihood (PPML) estimation  
>>> can be
>>> obtained with the command
>>>
>>> 	poisson depvar_ij ln(indepvar1_ij) ln(indepvar2_ij) ...
>>> ln(indepvarN_ij), robust
>>>
>>> I looked up Stata help for the command, to understand syntax and  
>>> such:
>>>
>>> 	www.stata.com/help.cgi?poisson
>>>
>>> Which simply says that the command fits a Poisson regression of  
>>> depvar
>>> on indepvars. However, in my google-searching, I noticed that  
>>> pseudo-
>>> maximum-likelihood estimation is sometimes called 'quasi-maximum,'  
>>> and
>>> that R has a "quasipoisson" family that seems to allow for
>>> overdispersion. So, am I missing something, or should I specify
>>> "quasipoisson" when implementing this estimation?
>>>
>>> Thanks a lot!
>>>
>>> Cheers,
>>>
>>>
>>> Wil
>>>
>>> ______________________________________________
>>> R-help <at> r-project.org mailing list
>>> https://stat.ethz.ch/mailman/listinfo/r-help
>>> PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
>>> and provide commented, minimal, self-contained, reproducible code.
>>
>>
>
> ______________________________________________
> R-help at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
> and provide commented, minimal, self-contained, reproducible code.